Assigning taxonomy to shotgun metagenomic reads: sequence alignment or composition?

[mutationalMeltdown]

I'm interested in identifying bacterial pathogens from shotgun metagenomic data. Crudely, I envisage the pipeline to be something like this:
1) QC reads
2) Filter out human sequences
3) Identify pathogen sequences

Typically it seems people use the following types of tools for the corresponding steps above:
1) Trimmomatic/cut-adapt
2) Bowtie2/SNAP/BWA against human reference genome(s)
3) Bowtie2/SNAP/BWA against reference bacterial genomes database

My question is can steps 2) and/or 3) be replaced with a sequence composition based approach? (E.g. a k-mer method, such as Kraken, LMAT or CLARK)

Of course k-mer methods will be faster but less sensitive than read alignment with a tool like Bowtie2. Additionally, read alignment uses paired-read and quality score info that k-mer approaches don't use. However, there seems to be an increasing trend to adopt 'alignment-free' or 'psudoalignment' approaches and I'm not sure how to evaluate the trade-offs. I'd be grateful for any advice. Thanks.

[fanli]

A recent analysis of HMP data used a k-mer approach (LMAT) to remove additional human variation:

http://genome.cshlp.org/content/25/7/1056.full

[mutationalMeltdown]

Yes I don't think LMAT is preferable to Kraken though; since it has a much larger memory footprint. However, this dosen't address the more fundamental question of whether k-mer approaches are preferable to read alignment or not